Extensive prior research has established that health disparities between race/ethnic and income groups are profoundly influenced by social determinants of health, such as poverty or access to healthy foods. While the existence of health disparities has been extensively chronicled in the research literature, far less research has evaluated the large-scale social programs-such as income support and food assistance programs-that address the social determinants of health, and are hypothesized to significantly reduce health disparities. The major scientific problem we seek to address is to identify whether, and to what degree, major social programs that affect the social determinants of health actually reduce health disparities. Answering this seemingly-simple question has been problematic due to persistent methodological limitations. Randomized trials of large-scale social programs have been prohibitively expensive, and are rarely conducted due to ethical dilemmas (e.g., randomizing hungry people to receive or not receive food support to identify their long-term type 2 diabetes risk violates obvious principles of beneficence). Ecological analyses conducted on aggregate data have suffered from ecological fallacies, while longitudinal cohort data have suffered from selection biases due to `confounding by indication'. Here, we propose a novel method called cohort filtering models that can evaluate the impact of large-scale social programs while overcoming the traditional barriers to unbiased evaluation. The method can be applied to a wide range of studies of social programs affecting health, crossing multiple domains of disease prevention programs, and crossing multiple domains of health behaviors and disparities outcomes. The strategy involves a novel nonparametric method to mimic a randomized controlled trial, combining the benefits of matching methods, instrumental variables techniques, and systems science simulations. The method will be applied to two longitudinal datasets we have assembled that allow us to link health and social program data at the individual level over several decades. We have linked health risk factor and health outcome data since 1968 and 1979, respectively, among two nationally-representative cohorts of Americans to administrative data revealing participation in eight major social support programs that reach over 1 in 3 Americans, ranging from income support to food assistance programs. Our preliminary research reveals how our unique approach can answer critical questions such as: how would altering the eligibility criteria for the country's largest nutrition programs be expected to alter disparities in obesity and type 2 diabetes? How much are race/ethnic disparities in preventable asthma exacerbations reduced by newly-expanded affordable housing programs? These are critical questions for health disparities research, which have not been possible to validly answer due to persistent methodological barriers to progress. In addition to testing and optimizing our new method, we will extend our user-friendly web-based data analysis interface using cloud-computing technologies so that any scientist can apply our method to their data or to public datasets, to test the effects of interventions on health disparities.